Agent rot / a pathology of autonomous systems

Your agent worked in March. It is quietly wrong in July. Nothing alerted.

MAR — OUTPUT QUALITY JUL — DASHBOARD STILL GREEN

agent rotnoun — the gradual, unattended decay of an AI agent's performance: stale context, drifted goals, and small compounding errors that turn yesterday's reliable automation into today's quiet liability.

Software has always rotted. The difference is that an agent keeps acting while it does.

Symptoms

Run the rot check — score one point per box

Green dashboards, declining output

Your metrics measure activity — tasks completed, tickets closed — not correctness. The agent stays busy while it gets worse.

Context older than the business

Prices changed, policies changed, people changed. Nobody re-fed the agent. It is executing confidently against a world that no longer exists.

Exceptions rising, readers gone

The exception queue creeps up, and the humans assigned to it stopped reading closely months ago because it was usually fine.

Prompt scar tissue

Layers of patched instructions nobody remembers the reason for. Each fix made sense once; together they contradict each other.

Vaguer handoffs

When the agent escalates to a human, the escalation carries less usable context than it did at launch. People re-do the work from scratch.

No last-evaluated date

Nobody can say when the agent's output was last checked against ground truth. Launch day does not count.

0–1: healthy. 2–3: early rot — treatable cheaply. 4+: the agent is a liability wearing an automation badge.

Causes

Five mechanisms, usually compounding

World drift. The environment the agent was tuned for keeps moving; the agent does not.

Context decay. Reference data, documents, and examples go stale silently — there is no expiry date on a prompt.

Feedback starvation. No loop compares output to ground truth, so errors never become signal.

Patch layering. Every incident adds instructions; nothing removes them. Complexity accretes until behavior is unpredictable.

Attention decay. Humans stop supervising what usually works. Reliability breeds the neglect that ends it.

Treatment

Rot is not curable. It is manageable — like maintenance, because it is maintenance

Rx1

Freshness SLAs on context

Every data source the agent relies on gets an expiry date and an owner. Stale context blocks the agent the way an expired certificate blocks a deploy.

Rx2

Scheduled re-evaluation

A recurring, calendared test of agent output against ground truth — sampled, scored, and trended. If quality is not measured on a schedule, it is not measured.

Rx3

Error budgets

Borrow from site reliability practice: an explicit tolerated error rate. Exceed it and the agent loses autonomy until it earns it back.

Rx4

Prune the prompt

Quarterly, delete every patched instruction that cannot justify itself. Scar tissue is removed, not managed.

Rx5

Rotate human attention

Randomize which outputs get deep human review, so no category of work goes permanently unwatched.

Agent rot is how it breaks. Bot blame is who answers when it does. The two failures travel together: unowned agents rot fastest.

About

Tarun Gulati has spent three decades leading commercial programs across North American banking, wealth and asset management — institutions where autonomous agents are now doing work that used to carry a person's name. He writes here about how those agents degrade, and at botblame.com about who is accountable when they do.

Conversation welcome: tarun.gulati@gmail.com